There are some good critiques of the IHME model in here IMO but this line bothered me. Model outputs *should* change in response to new data and when you're dealing with nonlinear systems, small changes can have a reasonably large effect. statnews.com/2020/04/17/inf… 
The death projections in their US model have shifted around within a range of ~60K to ~90K. That strikes me as quite reasonable, under the circumstances.
Other things have changed more, i.e. their projections on hospitalizations have decreased by 3x to 5x. That's not great.
Other things have changed more, i.e. their projections on hospitalizations have decreased by 3x to 5x. That's not great.
Another question is: should the structure/parameters/assumptions of the model change as you collect more data? In other words, not just the inputs, but the way you're processing those inputs? And maybe which inputs you're looking at? That's complicated.
My default answer is "sparingly; you should avoid it". But I'm working in fields (i.e. electoral politics) where the structures are well-established going in. I am perhaps more sympathetic to structural changes with something like COVID-19 where we're all learning in real time.
But unless you've built a model that *updates in real time*, I think it's easy to *underestimate* how much a well-designed model *should* change—certainly not all the time, but in response to high-leverage data. Most experts' intuitions are a bit miscalibrated here.
This can be defined mathematically: you want to design a model as to minimize serial correlation. Some experts prefer forecasts that change only gradually so they won't get criticized for changing their mind. But this intuition is often wrong. More here: fivethirtyeight.com/features/why-s…
